fix(workflows): make expression operator/literal parsing quote-aware

[jawwad-ali]

Description

The workflow expression evaluator (_evaluate_simple_expression) split on operator keywords with naive str.find/split, so a keyword inside a quoted operand was mistaken for an operator. Two concrete bugs:

1. Keyword inside a quoted string — {{ inputs.mode == 'read and write' }} split on the inner and and evaluated as (inputs.mode == 'read) and (write') -> wrong result.
2. Greedy literal short-circuit — 'a' == 'b' matched startswith("'") and endswith("'") and was stripped to the garbage truthy string a' == 'b. So {{ 'done' == 'failed' }} evaluated truthy and gated the wrong branch in if/switch/while/fan-out steps.

The docstring already lists comparisons, and/or/not, in/not in, and string literals as supported, so silent mis-parsing of common human-readable strings ("approve or reject", "read and write", "in progress") contradicts the contract.

Fix

• Add a quote/bracket-aware _find_top_level(text, token) helper — a direct sibling of the existing _split_top_level_commas (added in fix(workflows): preserve commas inside quoted list-literal elements #3134) — and use it for the or/and/comparison/in/not in splits, so operators inside quoted operands are skipped.
• Tighten the literal short-circuit to fire only when the opening quote's matching close is the final character (otherwise fall through to operator parsing).

Pipe-filter handling is untouched (it runs first and already worked for single filters).

Testing

• uvx ruff check clean.
• New TestExpressions::test_operator_splitting_is_quote_aware — fails before / passes after; covers keyword-in-literal (both operand sides), membership, literal-vs-literal equality, single-literal preservation, plus regression guards (plain comparisons, real and/or, | default('a and b')).
• Full tests/test_workflows.py: 305 passed, 1 skipped. (5 pre-existing symlink tests fail only on Windows due to the SeCreateSymbolicLink privilege — identical on clean main with this change stashed; they pass on CI Linux.)
• Verified each case against the real module on main (b7e67f5).

AI Disclosure

• I did use AI assistance (describe below)

Found and fixed with Claude Code (Claude Opus 4.8) under my direction. AI located the quote-naive splitting + greedy literal short-circuit, mirrored the existing _split_top_level_commas pattern, and drafted the regression tests; I reproduced every mis-parse against the real evaluator, confirmed the regression guards (filter args, real boolean operators, list literals) still pass, and reviewed the diff before submitting.